Inference and Docker¶

The inference interface classifies non-overlapping 10-second ECG windows. It does not train models or download datasets.

Select a model¶

Model	Leads	Classes
`12lead-conformer`	12	acceptable / unacceptable
`12lead-rbfsvm`	12	acceptable / unacceptable
`singlelead-conformer`	1	good / medium / bad
`singlelead-rbfsvm`	1	good / medium / bad

See the model catalogue for provenance and hashes.

Local command¶

python -m src.ecg_sqi_inference predict \
  --model 12lead-conformer \
  --input /path/to/input \
  --fs 500 \
  --out /path/to/output \
  --device cpu

--input may be one file or a recursively scanned directory. Supported inputs are .npy, .npz, numeric .csv, and WFDB .hea records. The source sampling frequency is supplied once with --fs; data are resampled to 125 Hz.

Shape contract¶

Single-lead: (samples,), (samples, 1), or (1, samples).
Twelve-lead: (samples, 12) or (12, samples).
At least 1,250 resampled samples are required.
An incomplete final window is reported as dropped_seconds and not padded.

Docker¶

Build once from the repository root:

docker build -f docker/inference/Dockerfile -t ecg-sqi-infer .
docker run --rm ecg-sqi-infer verify-bundles

Mount the same directory for input and output:

docker run --rm -v /host/ecg:/data ecg-sqi-infer predict \
  --model singlelead-conformer \
  --input /data/input \
  --fs 500 \
  --out /data/output

For WSL, Windows drives appear below /mnt; for example, E:\ecg-data becomes /mnt/e/ecg-data.

Python API¶

from pathlib import Path

from src.ecg_sqi_inference.core import predict_records
from src.ecg_sqi_inference.models import get_predictor

summary = predict_records(
    input_path=Path("example.npy"),
    out_dir=Path("example-output"),
    fs=125,
    predictor=get_predictor("singlelead-rbfsvm"),
)

The stable functions and error contracts are documented in the Python API.